Early prediction of extubation failure in AECOPD patients using emergency department data: a multicenter interpretable machine learning study

Purpose Weaning intubated patients with acute exacerbation of chronic obstructive pulmonary disease (AECOPD) from mechanical ventilation remains a major challenge in critical care. Current pre-extubation assessments often yield suboptimal accuracy for evaluating weaning readiness. Furthermore, while machine learning (ML) has shown promising predictive potential, existing models are largely limited by their opaque nature and insufficient rigorous external validation. We aimed to develop and externally validate a transparent ML model exclusively using early emergency department (ED) resuscitation data, to generate patient-specific weaning risk predictions and guide stratified intensive care unit (ICU) management. Methods This retrospective multicenter study enrolled 535 consecutive patients with AECOPD receiving mechanical ventilation at Yixing People’s Hospital (January 2018–January 2023). The cohort was randomly partitioned into a derivation set ( n = 375) and an internal validation set ( n = 160) at a 7:3 ratio using outcome-stratified sampling. A prospective observational cohort for temporal external validation ( n = 200) was subsequently established at Jiangsu Provincial People’s Hospital (May 2023–May 2025). Nineteen pre-intubation variables were collected. A dual-stage LASSO-Boruta pipeline was applied to identify robust predictors. Six machine learning (ML) algorithms were trained with class-weight adjustments to address outcome imbalance. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC), calibration curves, and Decision Curve Analysis (DCA), with the SHapley Additive exPlanations (SHAP) framework integrated for model interpretability. Results Seven core predictors were identified: oxygenation index (PaO 2 /FiO 2 ), left ventricular ejection fraction (LVEF), white blood cell (WBC) count, baseline lactate, diaphragmatic excursion, chronological age, and serum creatinine. The Random Forest (RF) model achieved optimal performance, with AUCs of 0.880 (95% CI 0.814–0.947) in the internal validation cohort and 0.874 (95% CI 0.803–0.945) in the external validation cohort.The model showed robust calibration, and DCA confirmed favorable net clinical benefit across clinically relevant risk thresholds. SHAP analysis revealed that early baseline oxygenation status, intrinsic cardiac reserve (LVEF), and systemic inflammatory burden were the primary drivers of prediction. A user-friendly web application was developed for real-time risk stratification. Conclusion By synthesizing multimodal signals obtained in the early phase of ED resuscitation, this SHAP-enhanced RF classifier provides a transparent and generalizable tool for predicting extubation outcomes in AECOPD patients. This approach enables upstream identification of dynamic weaning failure risk, allowing intensivists to optimize personalized weaning strategies. Clinical trial number Not applicable.